Minimal implementation to solve the HalfCheetah env using open-loop oscillators

halfcheetah_minimal.py

import gymnasium as gym
import numpy as np
from gymnasium.envs.mujoco.mujoco_env import MujocoEnv

# Env initialization
env = gym.make("HalfCheetah-v4", render_mode="human")
# Wrap to have reward statistics
env = gym.wrappers.RecordEpisodeStatistics(env)
mujoco_env = env.unwrapped
n_joints = 6
assert isinstance(mujoco_env, MujocoEnv)
# PD Controller gains
kp, kd = 1.0, 0.05
# Reset the environment
t, _ = 0.0, env.reset(seed=0)
# Oscillators parameters
omega_stance = 2 * np.pi * 4.622 * np.ones(n_joints)
omega_swing = 2 * np.pi * 3.865 * np.ones(n_joints)
phase_shifts = 2 * np.pi * np.array([0.00, 0.789, 0.316, 0.294, 0.629, 0.921])
amplitudes = np.array([1.123, -1.91, -1.204, 1.173, 1.196, -0.085])
offsets = np.array([-0.114, 0.075, 0.002, -0.493, -0.501, -0.227])
oscillator_dt = 0.001  # 1kHz, integration step
# Initial joint positions
theta = phase_shifts.copy()

while True:
    env.render()
    # Integrate oscillators equations
    for _ in range(int(mujoco_env.dt / oscillator_dt)):
        in_swing_phase = np.sin(theta) > 0
        theta_dot = in_swing_phase * omega_swing + (1 - in_swing_phase) * omega_stance
        # Integrate and keep theta in [0, 2 * pi]
        theta = (theta + oscillator_dt * theta_dot) % (2 * np.pi)

    # Open-Loop Control using oscillators
    desired_qpos = amplitudes * np.sin(theta) + offsets

    # PD Control: desired qvel is zero
    desired_torques = (
        kp * (desired_qpos - mujoco_env.data.qpos[-n_joints:])
        - kd * mujoco_env.data.qvel[-n_joints:]
    )
    desired_torques = np.clip(desired_torques, -1.0, 1.0)  # clip to action bounds
    _, reward, terminated, truncated, info = env.step(desired_torques)
    t += mujoco_env.dt

    if terminated or truncated:
        print(f"Episode return: {float(info['episode']['r'].item()):.2f}")
        t, _ = 0.0, env.reset()
        # Reinitialize
        theta = phase_shifts.copy()

Paper: https://arxiv.org/abs/2310.05808